Database reduction method according to phoneme distribution of large text corpus and usage in corpus-based speech synthesis

摘要

PURPOSE: A method for reducing a synthesis database depending on distribution and frequency of each phoneme is provided to keep diverse tones/rhythms while removing an unnecessary unit by reflecting the distribution of each synthesis unit and the frequency in synthesis on a synthesizer using a large capacity synthesis database. CONSTITUTION: A targeted reduction number of each synthesis unit is fixed by analyzing the distribution of each synthesis unit with a large capacity text(201). Synthesizing the large capacity text, a voice synthesizer records a selected segment number and accumulates the frequency of each segment number(203). Representative values are selected/clustered while being merged by using the frequency, and the characteristic values of the tone/rhythm and phonological environment(205). The synthesis database(14) is constructed by the clustered final representative values.